# What Thousands of Steel-Plant Drone Frames Reveal About the Camera and AI You Actually Need

> What a drone camera can and cannot see on a steel plant, how to test it, and what the camera, lens and data link have to deliver.

Source: https://beyondpixels.dev/blog/steel-plant-drone-frames/ · Published 2026-10-15 · Evidence: measured, derived, published, illustrative

Interactive instruments in the web version are shown here as their component tags with default parameters.

<Summary title="The short version">

- A close-up in an inspection report is usually a crop of the wide photo. Detail the camera did not record cannot be added back.
- Work backwards from the defect. At 17 m, a 20 MP camera with 1.4 µm pixels and a 6 mm lens records about 4 mm per pixel.
- At that size, lifted and missing sheets are easy to see. A 10 mm hole is at the edge of what can be spotted, and cracks cannot be measured.
- Many serious defects sit on walls, corners and gable ends, so fly side-on views as well as the roof.
- Light ruins more frames than resolution does: backlight, deep shade and dust.
- Motion, lens sharpness and the data link all take pixels away or add cost, and each can be calculated before the first flight.
- To know how often a camera finds a defect, plant known defects on a test wall and count.

</Summary>

Open almost any drone inspection report and you will find the same page: a wide photo with a red box, a zoomed close-up beside it, a severity rating and a recommendation.

<Plate kind="diagram" src="/media/steel-plant-drone-frames/report-page.webp" alt="A typical inspection report row: wide drone view with a red box, a zoomed close-up of the same frame, the category, a severity letter and a recommendation" caption="A typical report row. The close-up on the right is usually a crop of the wide frame, not a second photo." evidence="illustrative" width="1600" height="800" />

That close-up is rarely a second photo taken from nearer. It is the same photo, enlarged on screen, so whatever the camera did not record the first time is gone for good.

<p class="pull">The pixel count is the most zoom an inspector will ever have.</p>

So the useful question is what the camera has to record for a finding to hold up.

## Start from the defect

The first number to know is how many millimetres of the structure one pixel covers. Surveyors call this the ground sample distance, or GSD.

Two details matter. The distance is the straight line from the lens to the surface, not the drone's height: a drone 110 m up beside a chimney and 20 m from the steel is working at 20 m. And the lens counts as much as the sensor, because the same camera with a wider lens sees more area in less detail.

<p class="formula">GSD (mm/px) = distance × pixel size ÷ focal length</p>

Our worked example is a 20 MP sensor with 1.4 µm pixels and a 6 mm lens, 17 m from a wall. It gives 3.97 mm per pixel, which we round to 4 mm and use throughout. The sensor is the onsemi AR2020, the one e-con Systems fits in its [e-CAM200](https://www.e-consystems.com/camera-modules/20mp-ar2020-high-resolution-camera-module.asp) module. The pixel budget card further down runs the same sum for any camera, lens and distance.

<EconCameras ids={["ecam200"]} compact />

<Plate kind="diagram" src="/media/steel-plant-drone-frames/gsd-worked-example.webp" alt="Similar-triangles diagram: a 20 MP sensor with 1.4 µm pixels and a 6 mm lens at 17 m from a wall; each pixel covers about 4 mm" caption="Similar triangles: one 1.4 µm pixel behind a 6 mm lens covers 3.97 mm of wall at 17 m." evidence="derived" meta="20 MP · 1.4 µm · f = 6 mm · R = 17 m" width="1600" height="900" />

## The emergencies are on the walls

A neat top-down grid is the natural flight plan, the way a mapping survey works. But many of the defects that matter on industrial buildings are on the sides. Wind suction is strongest at roof edges and corners, which is where cladding and flashing work loose, and you cannot see a wall from straight above.

<BarChart title="Lifted and uprooted sheets dominate the emergencies" unit="%" evidence="measured" caption="Share of CAT-A (emergency) findings that involve each defect type in the analysed missions. The shares add to more than 100 % because a finding can involve more than one type." rows={[{label: "Uprooted or lifted sheets", value: 45}, {label: "Damaged sheets", value: 25}, {label: "Loose material on roofs", value: 20}, {label: "Perforations", value: 9}, {label: "Missing sheets", value: 5}, {label: "Corrosion", value: 3}]} />

Vertical side cladding was named in several CAT-A entries, ahead of roof sheets, flashing and louvres, and the locations cluster at gable ends, corners and around ducts. Dust and material build-up was the most common finding overall at 36 %, but only 9 % of those were CAT-A.

<Plate kind="diagram" src="/media/steel-plant-drone-frames/findings-by-severity.webp" alt="Stacked bar chart of drone inspection findings by type and severity, from dust and material build-up at the top to handrail and walkway at the bottom" caption="What drone inspectors flag, split by severity from CAT-A (act now) to keep under observation." evidence="measured" meta="Inspection findings · all missions" width="1536" height="1024" />

The flight logs agree. The median camera tilt was −30°, only 13 % of frames were close to straight down, and on stack missions 11 % looked upward.

<Plate kind="diagram" src="/media/steel-plant-drone-frames/camera-tilt.webp" alt="Three drone sketches: median camera tilt −30°, 13% near-nadir frames, 11% upward-looking frames on stack missions" caption="Where the camera actually pointed, from flight metadata." evidence="measured" meta="Gimbal pitch · flight logs" width="1600" height="800" />

<p class="pull">The findings that matter are side-on, at corners and in mixed light, so the camera, gimbal and flight plan should be built for those views.</p>

In practice that means a gimbal that tilts above the horizon, lenses chosen for 10 to 30 m stand-off instead of 100 m altitude, and a flight pattern that shoots every gable end and corner from two sides.

<Plate kind="diagram" src="/media/steel-plant-drone-frames/mission-pattern.webp" alt="Mission design: side-on views at corners, gimbal travel above the horizon, 10 to 30 m stand-off, and an orbit that shoots every gable end and corner from two sides" caption="Designing the mission around side-on findings." evidence="illustrative" width="1600" height="800" />

<Plate src="/media/steel-plant-drone-frames/gable-cladding.webp" alt="Weathered blue and rust-coloured cladding on a raised industrial building, with a sheet lifting at the gable end" caption="Gable ends and corners are where cladding works loose." evidence="illustrative" meta="AI-generated illustration" width="1536" height="1024" />

## What 4 mm per pixel can and cannot show

Imaging science has a long-standing rule of thumb, the Johnson criteria. To have even odds of completing a task, you need about 2 pixels across the narrowest side of the target to spot it, 8 to classify it and 13 to identify it. The rule was fitted to people looking at vehicles, so treat it as a guide for planning. For a more cautious plan, move to the 90 % level, which needs about 3.5, 14 and 23 pixels.

Hands-on visual inspection is usually done within about 600 mm, where normal eyesight resolves about 0.17 mm. Our drone camera records 4 mm per pixel, about 23 times coarser. To match the eye at 600 mm you would have to fly within 0.73 m of the surface, or 1.5 m with a 12 mm lens. Nobody does that along a conveyor gallery.

<Plate kind="diagram" src="/media/steel-plant-drone-frames/eye-vs-drone.webp" alt="Human eye at 600 mm resolves 0.17 mm; drone camera at 17 m records 4 mm per pixel" caption="The eye at 600 mm against the drone camera at 17 m." evidence="derived" meta="1 arc-minute · 20 MP, 1.4 µm, about 6 mm" width="1600" height="800" />

Sheet-level defects, which drive most CAT-A findings, are well inside the limit. A missing sheet covers hundreds of pixels.

<Plate src="/media/steel-plant-drone-frames/missing-sheet.webp" alt="A missing cladding sheet leaves a dark rectangular opening in a corrugated wall" caption="A missing sheet spans hundreds of pixels at 4 mm per pixel." evidence="illustrative" meta="AI-generated illustration" width="1536" height="1024" />

Holes are a different story. A 10 mm perforation covers about 2.5 pixels, which is at the edge of what can be spotted. A 60 mm hole covers about 15, enough to classify. A 0.3 mm crack is a fraction of one pixel wide, so at any practical stand-off it cannot be measured or classified. A long, high-contrast line can sometimes still be spotted, and then the drone flags the area and a technician measures it.

The card below shows how many pixels a detail gets. It starts with the e-CAM200 sensor at 17 m with a 6 mm lens, looking at a 10 mm hole.

<PixelBudget pitchUm={1.4} widthPx={5120} focalMm={6} rangeM={17} />

## Measure the probability of detection

Every inspection method has to say how likely it is to find a defect of a given size. This is the probability of detection, or POD, and drone camera datasheets rarely give it. You can measure it yourself.

1. Build a test wall from real sheeting with planted defects: lifted edges, holes from 10 to 60 mm, missing fasteners, a bent rail.
2. Fly it at your planned distance and speed, with the production camera, lens and image settings.
3. Ask inspectors to mark what they find, without telling them where the defects are.
4. Plot hits and misses against defect size. MIL-HDBK-1823A describes the standard way to do this. Use dozens of planted defects, not a handful.
5. Change one thing at a time (lens, distance, shutter speed, compression) and see what improves detection.

<Plate kind="diagram" src="/media/steel-plant-drone-frames/pod-test-wall.webp" alt="POD workflow: a test wall with planted defects, a drone flying the production setup, blinded inspectors marking findings, and a POD curve against defect size" caption="The five steps of a POD trial." evidence="illustrative" width="1600" height="800" />

The result tells you whether a better camera really moves the a90/95, the defect size found 90 % of the time with 95 % confidence, and it gives you a number to show an asset owner.

The size of the test matters. Finding 29 out of 29 planted defects of one size is enough to show a detection rate of at least 90 % at 95 % confidence. A trial planned to allow one miss needs 46 targets, and two misses need 61.

## Light ruins more frames than resolution

The failures that never appear on a datasheet are about light.

Backlight is the first. Looking up at a chimney top, the camera exposed for the sky at 1/1600 s, f/5 and ISO 100, about EV 15.3, and the steel went almost black. Brightening it afterwards brings up noise and banding, not rust texture.

Shadow is the second. Gutters under the eaves, louvre interiors and the undersides of decks sit in deep shade next to sunlit roof, and a choked gutter you cannot see is a CAT-A you did not log.

Dust is the third. Plant dust turns everything pinkish grey. Auto white balance corrects for it differently from frame to frame, which destroys the colour that tells fresh rust from old paint.

<Plate src="/media/steel-plant-drone-frames/stack-backlit-vs-exposed.webp" alt="Two views of a steel stack: on the left exposed for the bright sky so the structure is a black silhouette, on the right exposed for the steel so rust is visible" caption="Exposed for the sky (left) and for the steel (right)." evidence="illustrative" meta="AI-generated illustration of a measured condition (≈ EV 15.3)" width="1536" height="1024" />

<Plate src="/media/steel-plant-drone-frames/stack-dust-haze.webp" alt="A stack top seen through pinkish-grey plant dust haze" caption="Plant dust shifts colour, and auto white balance shifts it differently every frame." evidence="illustrative" meta="AI-generated illustration" width="1536" height="1024" />

Three things help. The first is on-sensor HDR and larger pixels for backlit and shaded work. A 2.0 µm pixel collects about twice the light of a 1.4 µm one, and the IMX678 in e-con Systems' [e-CAM810](https://www.e-consystems.com/camera-modules/4k-sony-starvis2-imx678-low-light-camera-module.asp) adds Clear HDR and DOL HDR modes.

<EconCameras ids={["ecam810"]} compact />

The second is white balance locked for the whole mission, set against a grey card at take-off, together with a tuned image pipeline. e-con Systems tunes colour for its cameras with its [TintE ISP](https://www.e-consystems.com/vision-ip-cores.asp). The third is RAW 10 or 12 bit, or high-quality encoding. Dust surveys are a low-contrast problem, and heavy compression erases the texture that shows the ribs are still visible and the layer is thin.

## Yearly inspections need a camera that does not drift

Most structures are inspected again on a schedule, often every year. The camera then becomes an instrument for measuring change, and an instrument must not drift.

- Fix the focus for the real inspection range instead of letting autofocus hunt across plain sheeting, so there is no focus difference between this year and next.
- Keep the colour pipeline the same from mission to mission: same image tuning, same white balance method, same encoding settings.
- Record how each photo was taken: camera angle, height, time and, if possible, the distance to the surface. Next year's photo can then be matched to this year's.
- If you run a fleet, ask the vendor how colour and sharpness are held consistent across production units. That matters more than a headline spec.

<Plate kind="diagram" src="/media/steel-plant-drone-frames/repeatability.webp" alt="Five practices for repeatable inspections: fixed focus, locked colour pipeline, same image settings, recorded capture geometry, unit-to-unit consistency" caption="What keeps this year's frame comparable with next year's." evidence="illustrative" width="1600" height="800" />

<Plate src="/media/steel-plant-drone-frames/corroded-cladding.webp" alt="Close view of corroded corrugated cladding with streaks of rust" caption="Corrosion is tracked as change over time, which only works if the camera does not drift." evidence="illustrative" meta="AI-generated illustration" width="1536" height="1024" />

## Motion, vibration and the shutter

A moving drone blurs the image, and the blur depends on speed and exposure time. The arithmetic is short. At 4 mm per pixel, a drone drifting at 1 m/s during a 1/240 s exposure moves the image by about 4 mm, one pixel. Flying along a gallery at 2 m/s needs 1/1000 s to keep the blur to half a pixel. Gimbal jitter of 1° per second at 1/240 s adds about 0.3 pixel with a 6 mm lens, and it grows with focal length, so it gets worse when you fit a longer lens for more detail.

The card below does these sums for your own settings. It starts at 2 m/s and 1/1000 s, with the 6 mm lens at 17 m.

<MotionBudget speed={2} shutter={1000} jitter={1} />

Sharpness also depends on the lens and the aperture. Diffraction sets a hard limit: with 1.4 µm pixels at f/5.6, contrast falls to zero at 0.455 cycles per pixel, just below the 0.5 the sensor can sample, so the finest tenth of the band is already empty before any lens flaw. Focus errors cost contrast in the same way.

The next card adds up every loss between the wall and the sensor. It starts at f/5.6 with 1.4 µm pixels, a 6 mm lens and a 17 m stand-off.

<SharpnessBudget pitchUm={1.4} focalMm={6} aperture={5.6} rangeM={17} speed={2} shutter={1000} />

The right shutter type depends on the camera's job.

- For inspection photos, a high-resolution rolling-shutter sensor works well while the drone hovers.
- For navigation and obstacle cameras, use a global shutter. A rolling shutter reads the image line by line, so fast movement bends straight edges, and that confuses navigation software (see rolling shutter). Stereo pairs also have to capture at the same moment. Two unsynchronised cameras at 60 fps can be up to one frame apart, 16.7 ms, and that is enough to create false depth when the drone is moving.

Two e-con Systems modules fit this job: the [e-CAM217](https://www.e-consystems.com/camera-modules/ar0234-global-shutter-camera-module.asp) (AR0234, 1920 × 1200 at 60 fps) for plain navigation and the [e-CAM521](https://www.e-consystems.com/camera-modules/5mp-sony-pregius-imx568-global-shutter-camera-module.asp) (IMX568, 5 MP) when stereo or navigation needs more resolution.

<EconCameras ids={["ecam217","ecam521"]} compact />

## Heat, steel and interference

- Heat: cast houses, stacks and hot-gas ducts raise the local air temperature well above ambient. Check the exact rating against your flight conditions and give the camera airflow in the payload design.
- Steel and compasses: large steel structures distort the magnetic field a drone's compass relies on, and tall structures can block satellite signals. That is exactly where you inspect, and it is why navigation from cameras is worth having.
- Electrical interference: fast camera links sit next to radios and GPS. Ask for EMC test documents for the exact camera model, and test the complete payload yourself.
- Supply chain: if your programme has NDAA requirements, ask for a declaration for the specific model, not a general claim.

<Plate kind="diagram" src="/media/steel-plant-drone-frames/heat-steel-interference.webp" alt="Four payload risks: heat near stacks, steel structures disturbing compass and GPS, electrical interference from camera links, and supply-chain declarations" caption="The payload risks that show up at inspection range." evidence="illustrative" width="1600" height="800" />

## Plan the data before the drone

A serious inspection drone often carries several cameras, each with its own job. This example rig uses e-con Systems cameras, and the plate shows the raw data it produces.

<Plate kind="diagram" src="/media/steel-plant-drone-frames/data-path-rig.webp" alt="A five-camera inspection drone: 20 MP inspection camera at 5.90 Gbit/s, 4K HDR camera at 2.99 Gbit/s, a navigation pair at 2.76 Gbit/s and a landing camera at 1.11 Gbit/s, 12.75 Gbit/s in total" caption="Five cameras, 12.75 Gbit/s of raw data." evidence="derived" meta="Raw rates · width × height × bits × fps" width="1600" height="800" />

Three numbers are worth bringing to the first design meeting.

- 12.75 Gbit/s decides how you connect the cameras and which computer you need on board.
- 24.6 MB is one 20 MP RAW photo at 10 bit. One photo every 2 seconds on a 20-minute flight comes to about 14.7 GB, so decide early what you keep.
- 80 tiles is what it takes to run defect detection on the drone without shrinking the photo. Squeezed to a 640-pixel-wide model input, 4 mm per pixel becomes 32 mm per pixel and small defects vanish. Cutting the photo into 640-pixel tiles with 20 % overlap takes about 80 tiles and keeps the detail.

Change the cameras, the flight or the model input size to see how the numbers move. The card starts with the five-camera rig above, with the e-CAM200 as the inspection camera.

<DataPath minutes={20} />

Cable length matters too. Standard camera ribbon cables are only 15 to 30 cm long. For cameras on arms or gimbals, use GMSL2, a camera link that reaches further.

## The camera was not the only bottleneck

On industrial sites, inspection days are lost to permits, weather, batteries and interference far more often than to the camera. A camera cannot shorten a permit queue. What it can do is raise the chance that one 20-minute flight produces usable evidence, because a re-flight costs another permit and another weather window.

## Specifying the camera

Image quality comes from the sensor, lens, image processing, cable and firmware working together, and that is how e-con Systems designs its cameras. Around the sensors above it offers M12 lens options from wide coverage to narrow detail, colour tuning for your site, MIPI CSI-2, GMSL2, USB and 10G Holoscan Sensor Bridge connections depending on the model, custom firmware (for example a global reset release on the AR2020 for flash-synchronised photos), and support for NVIDIA Jetson Orin NX, Nano and AGX Orin and Qualcomm platforms. Availability depends on the model, so check the [product pages](https://www.e-consystems.com/default.asp).

<Plate kind="diagram" src="/media/steel-plant-drone-frames/specify-the-camera.webp" alt="Six parts of a camera specification: sensor, lens, image processing, connection, firmware and computer" caption="Image quality comes from six parts working together." evidence="illustrative" width="1600" height="800" />

## Questions to ask a drone camera vendor

1. What GSD do I get at my real stand-off with the lens you propose?
2. Which lenses can you supply for 10 m, 17 m and 30 m stand-off, and what is their MTF at the sensor's Nyquist frequency?
3. Which HDR modes does the camera expose on my computing platform, and at what frame rates?
4. Can I lock white balance, exposure and image settings for a whole mission, and can I export RAW?
5. Which sensors are global shutter, and can they be synchronised in hardware?
6. What is the readout time in the mode I will use, and what does it mean for motion distortion?
7. What cable lengths do you support, and is GMSL2 available?
8. What is the operating temperature range of the exact model I am buying?
9. What are the power draw and weight with the lens fitted?
10. Can you provide EMC, environmental and NDAA documentation for the exact model?
11. How do you keep colour, focus and sharpness consistent across cameras in a fleet?
12. Will you help me measure POD on a seeded test wall, at my real stand-off, speed and lighting?

## Making training data when real defects are rare

If you are building AI to find defects, you will hit a familiar problem. Real defect photos are rare, they belong to the asset owner, and they are unevenly spread. There are thousands of photos of sound walls and very few of torn sheets.

We built a casebook of 16 inspection scenarios, from loose cladding to backlit chimney tops. Each one fixes the camera angle, stand-off, defect size, lighting and the visual limits that make an image behave like a real inspection frame. Drawing a torn sheet is not enough. The distance, perspective, light, material and surroundings have to be controlled as well. A training image for a torn conveyor-gallery sheet, for example, might be specified like this:

<p class="spec-line">17 m stand-off · 35° camera tilt · ~400 mm tear · low-angle light · corrugated steel · handrail and walkway in frame</p>

<Note kind="method" title="About the images in this article">

The plant photographs here are AI-generated illustrations made from the same kind of scenario description, and they are labelled as such. The statistics and charts come from the inspection data. The formulas and tools are derived from the camera parameters stated in the text.

</Note>

<div class="cta"><p><strong>Want the full defect casebook?</strong> The 16 scenarios, with their prompts and constraints.</p><a class="btn btn-primary" href="mailto:hello@beyondpixels.dev?subject=CASEBOOK">Request the casebook</a></div>